Special Topic: Missing Values

Missing Values Common in Real Data  Pneumonia: –6.3% of attribute values are missing –one attribute is missing in 61% of cases  C-Section: –only about 1/2% of attribute values are missing –but 27.9% of cases have at least 1 missing value  UCI machine learning repository: –31 of 68 data sets reported to have missing values

“Missing” Can Mean Many Things  Randomly missing: –usually best case –usually not true  Non-randomly missing  Presumed normal, so not measured  Causally missing: –attribute value is missing because of other attribute values (or because of the outcome value!)

Dealing With Missing Data  Throw away cases with missing values –in some data sets, most cases get thrown away –if missing not random, throwing away cases can bias sample towards certain kinds of cases  Treat “missing” as a new attribute value –what value should we use to code for missing with continuous or ordinal attributes? –if missing causally related to what is being predicted?

Dealing With Missing Values  Marginalize over missing values –Some learning methods handle missing data –Most don’t (including neural nets)  Impute (fill-in) missing values –once filled in, data set is easy to use –if missing values poorly predicted, may hurt performance of subsequent uses of data set

Imputing Missing Values  Fill-in with mean, median, or most common value  Predict missing values using machine learning  Expectation Minimization (EM): Build model of data values (ignore missing vals) Use model to estimate missing values Build new model of data values (including estimated values from previous step) Use new model to re-estimate missing values Re-estimate model Repeat until convergence

Potential Problems  Imputed values may be inappropriate: –in medical databases, if missing values not imputed separately for male and female patients, may end up with male patients with 1.3 prior pregnancies, and female patients with low sperm counts –many of these situations will not be so obvious  If some attributes are difficult to predict, filled-in values may be random (or worse)  Some of the best performing machine learning methods are impractical to use for filling in missing values (neural nets)

Research in Handling Missing Values  Lazy learning: –don’t train a model until you know test case –missing in test case may “shadow” missing values in train set  Better algorithms: –Expectation maximization (EM) –Non-parametric methods (since parametric methods often work poorly when assumptions are violated)  Faster Algorithms: –apply to very large datasets

Special Topic: Feature Selection

Anti-Motivation  Most learning methods implicitly do feature selection: –decision trees: use info gain or gain ratio to decide what attributes to use as tests. many features don’t get used. –neural nets: backprop learns strong connections to some inputs, and near-zero connections to other inputs. –kNN, MBL: weights in Weighted Euclidean Distance determine how important each feature is. weights near zero mean feature is not used. –Bayes nets: statistics in tables allow some features to have little or no effect on model.  So why do we need feature selection?

Motivation

Brute-Force Approach  Try all possible combinations of features  Given N features, 2 N subsets of features –usually too many to try –danger of overfitting  Train on train set, evaluate on test set (or use cross-validation)  Use set of features that performs best on test set(s)

Two Basic Approaches  Wrapper Methods: –give different sets of features to the learning algorithm and see which works better –algorithm dependent  Proxy Methods (relevance determination methods) –determine what features are important or not important for the prediction problem without knowing/using what learning algorithm will be employed –algorithm independent

Wrapper Methods  Wrapper methods find features that work best with some particular learning algorithm: –best features for kNN and neural nets may not be best features for decision trees –can eliminate features learning algorithm “has trouble with”  Forward stepwise selection  Backwards elimination  Bi-directional stepwise selection and elimination

Relevance Determination Methods  Rank features by information gain –Info Gain = reduction in entropy due to attribute  Try first 10, 20, 30, …, N features with learner  Evaluate on test set (or use cross validation)  May be only practical method if thousands of attributes

Advantages of Feature Selection  Improved accuracy!  Less complex models: –run faster –easier to understand, verify, explain  Feature selection points you to most important features  Don’t need to collect/process features not used in models

Limitations of Feature Selection  Given many features, feature selection can overfit –consider 10 relevant features, and 10 9 random irrelevant features  Wrapper methods require running base learning algorithm many times, which can be expensive!  Just because feature selection doesn’t select a feature, doesn’t mean that feature isn’t a strong predictor –redundant features  May throw away features domain experts want in model  Most feature selection methods are greedy and won’t find optimal feature set

Current Research in Feature Selection  Speeding-up feature selection (1000’s of features)  Preventing overfitting (1000’s of features)  Better proxy methods –would be nice to know what the good/relevant features are independent of the learning algorithm  Irrelevance detection: –truly irrelevant attributes can be ignored –better algorithms –better definition(s)

Bottom Line  Feature selection almost always improves accuracy on real problems  Plus: –simpler, more intelligible models –features selected can tell you about problem –less features to collect when using model in future Feature selection usually is a win.